Unveiling the Secrets without Data: Can Graph Neural Networks Be Exploited through Data-Free Model Extraction Attacks?

Graph neural networks (GNNs) play a crucial role in various graph applications, such as social science, biology, and molecular chemistry. Despite their popularity, GNNs are still vulnerable to intellectual property threats. Previous studies have demonstrated the susceptibility of GNN models to model extraction attacks, where attackers steal the functionality of GNNs by sending queries and obtaining model responses. However, existing model extraction attacks often assume that the attacker has access to specific information about the victim model's training data, including node attributes, connections, and the shadow dataset. This assumption is impractical in realworld scenarios. To address this issue, we propose STEAL-GNN, the first data-free model extraction attack framework against GNNs. STEALGNN advances prior GNN extraction attacks in three key aspects: 1) It is completely data-free, as it does not require actual node features or graph structures to extract GNN models. 2) It constitutes a full-rank attack that can be applied to node classification and link prediction tasks, posing significant intellectual property threats across a wide range of graph applications. 3) It can handle the most challenging hard-label attack setting, where the attacker possesses no knowledge about the target GNN model and can only obtain predicted labels through querying the victim model. Our experimental results on four benchmark graph datasets demonstrate the effectiveness of STEALGNN in attacking representative GNN models. learned different aspects of the victim model's decision criteria. By enforcing these inconsistencies and analyzing which applications trigger them, the attacker can gain insights into the victim model's complex risk assessment process.